Best agent harness for coding (2026): published scores, prices and limits

Updated 29 Oct 2026. No affiliate links: nobody pays us for placement on this page.

Harness vs model: the harness is the program around the model (Claude Code, Codex, Aider, OpenHands) that decides which files it reads, which tools it can run and how many retries it gets. Because of that, the same model scores differently in different harnesses, so a coding score only means something with both names attached.

For example, on 29 Oct 2026 Terminal-Bench 4.0 listed Claude Opus 5.5 at 64.8% in Claude Code, while other leaderboards using other harnesses list the same model between 59.6% and 65.15% (details on our Terminal-Bench page).

Top published result on each benchmark

One row per benchmark: the highest result we could trace to a primary leaderboard. The three benchmarks test different things, so do not compare scores across rows; this is not a ranking.

Notes: the official SWE-bench Verified board lists higher entries (e.g. 79.2%) that the SWE-bench team has not marked as checked, so we show the highest checked one. Its newest entry is from Feb 2026, and the benchmark is near saturation (see our SWE-bench page). Cursor has no entry on any of the three boards: not reported. Claude Code has no SWE-bench Verified or Aider Polyglot entry: not reported. Aider has no Terminal-Bench 4.0 entry: not reported. More on the third benchmark: Aider Polyglot explained.

Open-source harnesses

GitHub facts read from the GitHub API on 29 Oct 2026.

• Aider (Apache-2.0): Terminal pair-programmer that edits files and commits to git. About 49,800 GitHub stars; latest release v0.86.0 on 9 Aug 2025, last code push 22 May 2026, so check activity before you depend on it. See the related benchmark.
• OpenHands (MIT): Full autonomous agent with a sandboxed runtime, CLI and web UI (repo moved from All-Hands-AI to the OpenHands org). About 92,100 stars; release v1.34.0 on 27 Oct 2026. Scored 71.8% on SWE-bench Verified with GPT-5 (7 Aug 2025, checked run).
• Cline (Apache-2.0): VS Code extension agent that asks before each file edit or command; you bring your own API key. About 71,000 stars; release v4.1.26 on 26 Oct 2026.
• Codex CLI (Apache-2.0): OpenAI's terminal agent; the client is open source even though the models are not. About 131,500 stars; release rust-v0.165.0 on 27 Oct 2026.
• mini-SWE-agent (MIT): A deliberately tiny agent loop from the SWE-bench team, used as a neutral baseline harness. About 8,900 stars; release v2.4.6 on 23 Jul 2026. See the related benchmark.

Research prototype (not a product): DeepSeek dsh-superagent

Kept separate from the lists above because it is not a tool we recommend for production use. Source: DeepSeek's own dsh-superagent-prototype repository and README (MIT licence), read 2 Nov 2026.

• What it is: an open-source research prototype. The README says it is reference code, not an official DeepSeek product, may contain bugs, and is not identical to DeepSeek's internal version; some designs in DeepSeek's related article are not implemented.
• What it does, per the README: a coding agent for a single GitHub repository, packaged as a bundle for DeepSeek Harness. It listens to GitHub (and optionally Feishu); when an allowed person asks for work, it opens a work item, works in its own session and directory, and replies in the original conversation, for example with a pull request. A web panel shows what needs a person.
• Status: repository created 31 Oct 2026 (GitHub API, read 2 Nov 2026). No benchmark results are published for it, so it does not appear in the score table.

Pricing: Claude Code, Cursor and Codex

Each price below was checked 29 Oct 2026 on the vendor's own page; prices change, so confirm before you buy.

• Claude Code (claude.com/pricing, checked 29 Oct 2026): included from Pro at $20/month ($17/month billed yearly); Max from $100/month for 5x or 20x Pro usage; not on Free. The limit that bites: usage resets on a rolling 5-hour window, paid plans also have weekly caps, and Claude Code draws from the same pool as Claude chat.
• Cursor (cursor.com/pricing, checked 29 Oct 2026): Free; Plus $8/month; Premium $30; Super $100; Ultra $200; Teams from $30/user. The limit that bites: Plus covers Grok models only; Claude and GPT models are billed at API list price on top, and the included allowance refreshes weekly.
• Codex (developers.openai.com/codex/pricing, checked 29 Oct 2026): through ChatGPT Free, Go ($8), Plus ($20) and Pro ($100, $200 or $500). The limit that bites: on Plus, a 5-hour message budget (OpenAI estimates 15 to 160 local messages with GPT-6.1 Sol, 5 to 45 with GPT-6 Astra) plus weekly limits; Pro currently has no 5-hour cap. With an API key you pay API rates instead.

What users complain about

These are individual reports from public issue trackers, not verified facts, and are kept apart from the data above.

• 21 Jul 2026: Claude Code: a user upgraded to Max 20x but says weekly limits kept draining at the Max 5x rate (open issue). Read the thread
• 30 Jul 2026: Claude Code: report of a session limit being used up without any work being done (closed issue). Read the thread
• 18 Jun 2026: Codex on the Plus plan: users report the 5-hour budget emptying far faster after a change on 16 Jun 2026 (closed issue, 200+ comments). Read the thread
• 14 Aug 2026: Cline: tool output and file content get truncated in context, forcing large files to be read in chunks (open issue). Read the thread
• 23 Oct 2026: Aider Polyglot: a report that some JavaScript exercises still contain skipped tests, which would make those tasks easier than intended (open issue). Read the thread

How to choose

• Already paying for Claude: Claude Code, but plan around the 5-hour and weekly caps.
• Want an open-source client you can audit or self-host: Aider, OpenHands, Cline or Codex CLI, with your own API key.
• Comparing models fairly: hold the harness fixed (mini-SWE-agent is the common neutral choice) and change only the model.

Method

We read each benchmark's official leaderboard data on 29 Oct 2026 and took the top result the maintainers publish, noting who submitted it. GitHub numbers come from the GitHub API on the same day. Prices come from vendor pricing pages. We did not run any paid hands-on tests. We never rank across benchmarks, and empty cells say "not reported".

Benchmarking base models on academic tasks instead of coding agents? See EleutherAI's lm-eval toolkit.

Update log

• 2 Nov 2026: added a separate research-prototype note on DeepSeek's dsh-superagent, sourced from its repository README and licence.
• 29 Oct 2026: page published; all scores, GitHub numbers and prices checked that day.

Disclosure: no affiliate links on this page. Evals Review earns nothing if you sign up for any tool named here.